xfs_vfsops.c

linux 内核源代码

	base_lock_flags = XFS_ILOCK_SHARED;
	if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
		/*
		 * We need the I/O lock if we're going to call any of
		 * the flush/inval routines.
		 */
		base_lock_flags |= XFS_IOLOCK_SHARED;
	}

	XFS_MOUNT_ILOCK(mp);

	ip = mp->m_inodes;

	mount_locked = B_TRUE;
	vnode_refed  = B_FALSE;

	IPOINTER_CLR;

	do {
		ASSERT(ipointer_in == B_FALSE);
		ASSERT(vnode_refed == B_FALSE);

		lock_flags = base_lock_flags;

		/*
		 * There were no inodes in the list, just break out
		 * of the loop.
		 */
		if (ip == NULL) {
			break;
		}

		/*
		 * We found another sync thread marker - skip it
		 */
		if (ip->i_mount == NULL) {
			ip = ip->i_mnext;
			continue;
		}

		vp = XFS_ITOV_NULL(ip);

		/*
		 * If the vnode is gone then this is being torn down,
		 * call reclaim if it is flushed, else let regular flush
		 * code deal with it later in the loop.
		 */

		if (vp == NULL) {
			/* Skip ones already in reclaim */
			if (ip->i_flags & XFS_IRECLAIM) {
				ip = ip->i_mnext;
				continue;
			}
			if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0) {
				ip = ip->i_mnext;
			} else if ((xfs_ipincount(ip) == 0) &&
				    xfs_iflock_nowait(ip)) {
				IPOINTER_INSERT(ip, mp);

				xfs_finish_reclaim(ip, 1,
						XFS_IFLUSH_DELWRI_ELSE_ASYNC);

				XFS_MOUNT_ILOCK(mp);
				mount_locked = B_TRUE;
				IPOINTER_REMOVE(ip, mp);
			} else {
				xfs_iunlock(ip, XFS_ILOCK_EXCL);
				ip = ip->i_mnext;
			}
			continue;
		}

		if (VN_BAD(vp)) {
			ip = ip->i_mnext;
			continue;
		}

		if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
			XFS_MOUNT_IUNLOCK(mp);
			kmem_free(ipointer, sizeof(xfs_iptr_t));
			return 0;
		}

		/*
		 * If this is just vfs_sync() or pflushd() calling
		 * then we can skip inodes for which it looks like
		 * there is nothing to do.  Since we don't have the
		 * inode locked this is racy, but these are periodic
		 * calls so it doesn't matter.  For the others we want
		 * to know for sure, so we at least try to lock them.
		 */
		if (flags & SYNC_BDFLUSH) {
			if (((ip->i_itemp == NULL) ||
			     !(ip->i_itemp->ili_format.ilf_fields &
			       XFS_ILOG_ALL)) &&
			    (ip->i_update_core == 0)) {
				ip = ip->i_mnext;
				continue;
			}
		}

		/*
		 * Try to lock without sleeping.  We're out of order with
		 * the inode list lock here, so if we fail we need to drop
		 * the mount lock and try again.  If we're called from
		 * bdflush() here, then don't bother.
		 *
		 * The inode lock here actually coordinates with the
		 * almost spurious inode lock in xfs_ireclaim() to prevent
		 * the vnode we handle here without a reference from
		 * being freed while we reference it.  If we lock the inode
		 * while it's on the mount list here, then the spurious inode
		 * lock in xfs_ireclaim() after the inode is pulled from
		 * the mount list will sleep until we release it here.
		 * This keeps the vnode from being freed while we reference
		 * it.
		 */
		if (xfs_ilock_nowait(ip, lock_flags) == 0) {
			if ((flags & SYNC_BDFLUSH) || (vp == NULL)) {
				ip = ip->i_mnext;
				continue;
			}

			vp = vn_grab(vp);
			if (vp == NULL) {
				ip = ip->i_mnext;
				continue;
			}

			IPOINTER_INSERT(ip, mp);
			xfs_ilock(ip, lock_flags);

			ASSERT(vp == XFS_ITOV(ip));
			ASSERT(ip->i_mount == mp);

			vnode_refed = B_TRUE;
		}

		/* From here on in the loop we may have a marker record
		 * in the inode list.
		 */

		/*
		 * If we have to flush data or wait for I/O completion
		 * we need to drop the ilock that we currently hold.
		 * If we need to drop the lock, insert a marker if we
		 * have not already done so.
		 */
		if ((flags & (SYNC_CLOSE|SYNC_IOWAIT)) ||
		    ((flags & SYNC_DELWRI) && VN_DIRTY(vp))) {
			if (mount_locked) {
				IPOINTER_INSERT(ip, mp);
			}
			xfs_iunlock(ip, XFS_ILOCK_SHARED);

			if (flags & SYNC_CLOSE) {
				/* Shutdown case. Flush and invalidate. */
				if (XFS_FORCED_SHUTDOWN(mp))
					xfs_tosspages(ip, 0, -1,
							     FI_REMAPF);
				else
					error = xfs_flushinval_pages(ip,
							0, -1, FI_REMAPF);
			} else if ((flags & SYNC_DELWRI) && VN_DIRTY(vp)) {
				error = xfs_flush_pages(ip, 0,
							-1, fflag, FI_NONE);
			}

			/*
			 * When freezing, we need to wait ensure all I/O (including direct
			 * I/O) is complete to ensure no further data modification can take
			 * place after this point
			 */
			if (flags & SYNC_IOWAIT)
				vn_iowait(ip);

			xfs_ilock(ip, XFS_ILOCK_SHARED);
		}

		if (flags & SYNC_BDFLUSH) {
			if ((flags & SYNC_ATTR) &&
			    ((ip->i_update_core) ||
			     ((ip->i_itemp != NULL) &&
			      (ip->i_itemp->ili_format.ilf_fields != 0)))) {

				/* Insert marker and drop lock if not already
				 * done.
				 */
				if (mount_locked) {
					IPOINTER_INSERT(ip, mp);
				}

				/*
				 * We don't want the periodic flushing of the
				 * inodes by vfs_sync() to interfere with
				 * I/O to the file, especially read I/O
				 * where it is only the access time stamp
				 * that is being flushed out.  To prevent
				 * long periods where we have both inode
				 * locks held shared here while reading the
				 * inode's buffer in from disk, we drop the
				 * inode lock while reading in the inode
				 * buffer.  We have to release the buffer
				 * and reacquire the inode lock so that they
				 * are acquired in the proper order (inode
				 * locks first).  The buffer will go at the
				 * end of the lru chain, though, so we can
				 * expect it to still be there when we go
				 * for it again in xfs_iflush().
				 */
				if ((xfs_ipincount(ip) == 0) &&
				    xfs_iflock_nowait(ip)) {

					xfs_ifunlock(ip);
					xfs_iunlock(ip, XFS_ILOCK_SHARED);

					error = xfs_itobp(mp, NULL, ip,
							  &dip, &bp, 0, 0);
					if (!error) {
						xfs_buf_relse(bp);
					} else {
						/* Bailing out, remove the
						 * marker and free it.
						 */
						XFS_MOUNT_ILOCK(mp);
						IPOINTER_REMOVE(ip, mp);
						XFS_MOUNT_IUNLOCK(mp);

						ASSERT(!(lock_flags &
							XFS_IOLOCK_SHARED));

						kmem_free(ipointer,
							sizeof(xfs_iptr_t));
						return (0);
					}

					/*
					 * Since we dropped the inode lock,
					 * the inode may have been reclaimed.
					 * Therefore, we reacquire the mount
					 * lock and check to see if we were the
					 * inode reclaimed. If this happened
					 * then the ipointer marker will no
					 * longer point back at us. In this
					 * case, move ip along to the inode
					 * after the marker, remove the marker
					 * and continue.
					 */
					XFS_MOUNT_ILOCK(mp);
					mount_locked = B_TRUE;

					if (ip != ipointer->ip_mprev) {
						IPOINTER_REMOVE(ip, mp);

						ASSERT(!vnode_refed);
						ASSERT(!(lock_flags &
							XFS_IOLOCK_SHARED));
						continue;
					}

					ASSERT(ip->i_mount == mp);

					if (xfs_ilock_nowait(ip,
						    XFS_ILOCK_SHARED) == 0) {
						ASSERT(ip->i_mount == mp);
						/*
						 * We failed to reacquire
						 * the inode lock without
						 * sleeping, so just skip
						 * the inode for now.  We
						 * clear the ILOCK bit from
						 * the lock_flags so that we
						 * won't try to drop a lock
						 * we don't hold below.
						 */
						lock_flags &= ~XFS_ILOCK_SHARED;
						IPOINTER_REMOVE(ip_next, mp);
					} else if ((xfs_ipincount(ip) == 0) &&
						   xfs_iflock_nowait(ip)) {
						ASSERT(ip->i_mount == mp);
						/*
						 * Since this is vfs_sync()
						 * calling we only flush the
						 * inode out if we can lock
						 * it without sleeping and
						 * it is not pinned.  Drop
						 * the mount lock here so
						 * that we don't hold it for
						 * too long. We already have
						 * a marker in the list here.
						 */
						XFS_MOUNT_IUNLOCK(mp);
						mount_locked = B_FALSE;
						error = xfs_iflush(ip,
							   XFS_IFLUSH_DELWRI);
					} else {
						ASSERT(ip->i_mount == mp);
						IPOINTER_REMOVE(ip_next, mp);
					}
				}

			}

		} else {
			if ((flags & SYNC_ATTR) &&
			    ((ip->i_update_core) ||
			     ((ip->i_itemp != NULL) &&
			      (ip->i_itemp->ili_format.ilf_fields != 0)))) {
				if (mount_locked) {
					IPOINTER_INSERT(ip, mp);
				}

				if (flags & SYNC_WAIT) {
					xfs_iflock(ip);
					error = xfs_iflush(ip,
							   XFS_IFLUSH_SYNC);
				} else {
					/*
					 * If we can't acquire the flush
					 * lock, then the inode is already
					 * being flushed so don't bother
					 * waiting.  If we can lock it then
					 * do a delwri flush so we can
					 * combine multiple inode flushes
					 * in each disk write.
					 */
					if (xfs_iflock_nowait(ip)) {
						error = xfs_iflush(ip,
							   XFS_IFLUSH_DELWRI);
					}
					else if (bypassed)
						(*bypassed)++;
				}
			}
		}

		if (lock_flags != 0) {
			xfs_iunlock(ip, lock_flags);
		}

		if (vnode_refed) {
			/*
			 * If we had to take a reference on the vnode
			 * above, then wait until after we've unlocked
			 * the inode to release the reference.  This is
			 * because we can be already holding the inode
			 * lock when VN_RELE() calls xfs_inactive().
			 *
			 * Make sure to drop the mount lock before calling
			 * VN_RELE() so that we don't trip over ourselves if
			 * we have to go for the mount lock again in the
			 * inactive code.
			 */
			if (mount_locked) {
				IPOINTER_INSERT(ip, mp);
			}

			VN_RELE(vp);

			vnode_refed = B_FALSE;
		}

		if (error) {
			last_error = error;
		}

		/*
		 * bail out if the filesystem is corrupted.
		 */
		if (error == EFSCORRUPTED)  {
			if (!mount_locked) {
				XFS_MOUNT_ILOCK(mp);
				IPOINTER_REMOVE(ip, mp);
			}
			XFS_MOUNT_IUNLOCK(mp);
			ASSERT(ipointer_in == B_FALSE);
			kmem_free(ipointer, sizeof(xfs_iptr_t));
			return XFS_ERROR(error);
		}

		/* Let other threads have a chance at the mount lock
		 * if we have looped many times without dropping the
		 * lock.
		 */
		if ((++preempt & XFS_PREEMPT_MASK) == 0) {
			if (mount_locked) {
				IPOINTER_INSERT(ip, mp);
			}
		}

		if (mount_locked == B_FALSE) {
			XFS_MOUNT_ILOCK(mp);
			mount_locked = B_TRUE;
			IPOINTER_REMOVE(ip, mp);
			continue;
		}

		ASSERT(ipointer_in == B_FALSE);
		ip = ip->i_mnext;

	} while (ip != mp->m_inodes);

	XFS_MOUNT_IUNLOCK(mp);

	ASSERT(ipointer_in == B_FALSE);

	kmem_free(ipointer, sizeof(xfs_iptr_t));
	return XFS_ERROR(last_error);
}

/*
 * xfs sync routine for internal use
 *
 * This routine supports all of the flags defined for the generic vfs_sync
 * interface as explained above under xfs_sync.
 *
 */
int
xfs_syncsub(
	xfs_mount_t	*mp,
	int		flags,
	int             *bypassed)
{
	int		error = 0;
	int		last_error = 0;
	uint		log_flags = XFS_LOG_FORCE;
	xfs_buf_t	*bp;
	xfs_buf_log_item_t	*bip;

	/*
	 * Sync out the log.  This ensures that the log is periodically
	 * flushed even if there is not enough activity to fill it up.
	 */
	if (flags & SYNC_WAIT)
		log_flags |= XFS_LOG_SYNC;

	xfs_log_force(mp, (xfs_lsn_t)0, log_flags);

	if (flags & (SYNC_ATTR|SYNC_DELWRI)) {
		if (flags & SYNC_BDFLUSH)
			xfs_finish_reclaim_all(mp, 1);
		else
			error = xfs_sync_inodes(mp, flags, bypassed);
	}

	/*
	 * Flushing out dirty data above probably generated more
	 * log activity, so if this isn't vfs_sync() then flush
	 * the log again.
	 */
	if (flags & SYNC_DELWRI) {
		xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
	}

	if (flags & SYNC_FSDATA) {
		/*
		 * If this is vfs_sync() then only sync the superblock
		 * if we can lock it without sleeping and it is not pinned.
		 */
		if (flags & SYNC_BDFLUSH) {
			bp = xfs_getsb(mp, XFS_BUF_TRYLOCK);
			if (bp != NULL) {
				bip = XFS_BUF_FSPRIVATE(bp,xfs_buf_log_item_t*);
				if ((bip != NULL) &&
				    xfs_buf_item_dirty(bip)) {
					if (!(XFS_BUF_ISPINNED(bp))) {
						XFS_BUF_ASYNC(bp);
						error = xfs_bwrite(mp, bp);
					} else {
						xfs_buf_relse(bp);
					}
				} else {
					xfs_buf_relse(bp);
				}
			}
		} else {
			bp = xfs_getsb(mp, 0);
			/*
			 * If the buffer is pinned then push on the log so
			 * we won't get stuck waiting in the write for
			 * someone, maybe ourselves, to flush the log.
			 * Even though we just pushed the log above, we
			 * did not have the superblock buffer locked at
			 * that point so it can become pinned in between
			 * there and here.
			 */
			if (XFS_BUF_ISPINNED(bp))
				xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
			if (flags & SYNC_WAIT)
				XFS_BUF_UNASYNC(bp);
			else
				XFS_BUF_ASYNC(bp);
			error = xfs_bwrite(mp, bp);
		}
		if (error) {
			last_error = error;
		}
	}

	/*
	 * If this is the periodic sync, then kick some entries out of
	 * the reference cache.  This ensures that idle entries are
	 * eventually kicked out of the cache.
	 */
	if (flags & SYNC_REFCACHE) {
		if (flags & SYNC_WAIT)
			xfs_refcache_purge_mp(mp);
		else
			xfs_refcache_purge_some(mp);
	}

	/*
	 * If asked, update the disk superblock with incore counter values if we
	 * are using non-persistent counters so that they don't get too far out
	 * of sync if we crash or get a forced shutdown. We don't want to force
	 * this to disk, just get a transaction into the iclogs....
	 */
	if (flags & SYNC_SUPER)
		xfs_log_sbcount(mp, 0);

	/*
	 * Now check to see if the log needs a "dummy" transaction.
	 */

	if (!(flags & SYNC_REMOUNT) && xfs_log_need_covered(mp)) {
		xfs_trans_t *tp;
		xfs_inode_t *ip;

		/*
		 * Put a dummy transaction in the log to tell
		 * recovery that all others are OK.